diff --git a/internal/lua/adapters/opencodego.lua b/internal/lua/adapters/opencodego.lua
index 98022aa..b3528ee 100644
--- a/internal/lua/adapters/opencodego.lua
+++ b/internal/lua/adapters/opencodego.lua
@@ -31,12 +31,20 @@ end
 
 function adapter.build_headers(meta)
     local ts = tostring(meta.timestamp or "")
+    local src = (meta.source and meta.source.name) or ""
     return {
         ["User-Agent"] = adapter.headers["User-Agent"],
         ["x-opencode-client"] = "cli",
         -- project 固定：同网关实例共享一个工作区身份
-        ["x-opencode-project"] = string.sub(sha256_hex("llmsproxy|" .. (meta.source and meta.source.name or "")), 1, 32),
-        ["x-opencode-session"] = rand_id("ses_", "session|" .. ts),
+        ["x-opencode-project"] = string.sub(sha256_hex("llmsproxy|" .. src), 1, 32),
+        -- session 必须**按源稳定**，不能每请求换：上游的前缀缓存在同一 session
+        -- 内才复用。实测（同一段 6032 token 提示词）：
+        --   固定 session  -> 第 2 次命中 5888/6032，cached_tokens=5888
+        --   每请求换 session -> 永远 0 命中
+        -- 原先用 meta.timestamp 派生，等于每请求都是新会话，缓存永远无效，
+        -- 上游也无法做会话亲和路由。
+        ["x-opencode-session"] = rand_id("ses_", "session|llmsproxy|" .. src),
+        -- request id 仍每请求唯一（它只是请求标识，不参与缓存键）
         ["x-opencode-request"] = rand_id("msg_", "request|" .. ts .. "|" .. tostring(meta.body or "")),
     }
 end
diff --git a/internal/lua/adapters/opencodezen.lua b/internal/lua/adapters/opencodezen.lua
index 7940a84..c836fb0 100644
--- a/internal/lua/adapters/opencodezen.lua
+++ b/internal/lua/adapters/opencodezen.lua
@@ -32,12 +32,20 @@ end
 
 function adapter.build_headers(meta)
     local ts = tostring(meta.timestamp or "")
+    local src = (meta.source and meta.source.name) or ""
     return {
         ["User-Agent"] = adapter.headers["User-Agent"],
         ["x-opencode-client"] = "cli",
         -- project 固定：同网关实例共享一个工作区身份
-        ["x-opencode-project"] = string.sub(sha256_hex("llmsproxy|" .. (meta.source and meta.source.name or "")), 1, 32),
-        ["x-opencode-session"] = rand_id("ses_", "session|" .. ts),
+        ["x-opencode-project"] = string.sub(sha256_hex("llmsproxy|" .. src), 1, 32),
+        -- session 必须**按源稳定**，不能每请求换：上游的前缀缓存在同一 session
+        -- 内才复用。实测（同一段 6032 token 提示词）：
+        --   固定 session  -> 第 2 次命中 5888/6032，cached_tokens=5888
+        --   每请求换 session -> 永远 0 命中
+        -- 原先用 meta.timestamp 派生，等于每请求都是新会话，缓存永远无效，
+        -- 上游也无法做会话亲和路由。
+        ["x-opencode-session"] = rand_id("ses_", "session|llmsproxy|" .. src),
+        -- request id 仍每请求唯一（它只是请求标识，不参与缓存键）
         ["x-opencode-request"] = rand_id("msg_", "request|" .. ts .. "|" .. tostring(meta.body or "")),
     }
 end
diff --git a/internal/lua/toolcall_preservation_test.go b/internal/lua/toolcall_preservation_test.go
index 31e18ad..33ae354 100644
--- a/internal/lua/toolcall_preservation_test.go
+++ b/internal/lua/toolcall_preservation_test.go
@@ -231,3 +231,53 @@ func TestOpenCodeGoVsZenReasoning(t *testing.T) {
 		}
 	}
 }
+
+// TestOpenCodeSessionIsStableForCache pins that the opencode client header set
+// keeps x-opencode-session STABLE per source while x-opencode-request stays
+// unique per call.
+//
+// Measured against the live OpenCode Go endpoint: the upstream prefix cache is
+// session-scoped. Replaying the same 6032-token prompt hits 5888 cached tokens
+// when the session id is fixed, and 0 when it changes per request. Deriving the
+// session from meta.timestamp (the previous behaviour) made every request a new
+// session, so the cache could never hit.
+func TestOpenCodeSessionIsStableForCache(t *testing.T) {
+	for _, name := range []string{"opencodego", "opencodezen"} {
+		vm := NewVM(freshAdapterDir(t))
+		if err := vm.Start(); err != nil {
+			t.Fatal(err)
+		}
+
+		metaA := map[string]interface{}{
+			"timestamp": 1000,
+			"body":      `{"messages":[{"role":"user","content":"one"}]}`,
+			"source":    map[string]interface{}{"name": "somesource", "meta": nil},
+		}
+		metaB := map[string]interface{}{
+			"timestamp": 2000, // different second, different body
+			"body":      `{"messages":[{"role":"user","content":"two"}]}`,
+			"source":    map[string]interface{}{"name": "somesource", "meta": nil},
+		}
+		h1, err := vm.BuildHeaders(name, metaA)
+		if err != nil {
+			t.Fatalf("%s headers A: %v", name, err)
+		}
+		h2, err := vm.BuildHeaders(name, metaB)
+		if err != nil {
+			t.Fatalf("%s headers B: %v", name, err)
+		}
+
+		if h1["x-opencode-session"] == "" {
+			t.Fatalf("%s: session header must be set", name)
+		}
+		if h1["x-opencode-session"] != h2["x-opencode-session"] {
+			t.Errorf("%s: x-opencode-session must be stable across requests (%q vs %q) — "+
+				"the upstream prefix cache is session-scoped, a rotating session kills every cache hit",
+				name, h1["x-opencode-session"], h2["x-opencode-session"])
+		}
+		if h1["x-opencode-request"] == h2["x-opencode-request"] {
+			t.Errorf("%s: x-opencode-request must differ per request", name)
+		}
+		vm.Stop()
+	}
+}